Algorithm to model gene expression on Affymetrix chips without the use of MM cells

DNA microarrays are powerful tools for quantifying gene expression patterns. However, obtaining reliable estimates of gene expression from raw measurements on microarrays presents several problems due to background contributions, nonspecific probe response, possible variation in probe sensitivities, and possible nonlinear responses of the probes to transcript concentration. In an effort to address the nonspecific response of probes, Affymetrix GeneChip arrays use two probes for each measurement. One of these probes, the mismatch (MM) probe, is intended to reflect the nonspecific response of the corresponding perfect match (PM) probe. However, the reliability of this approach has not been established by published experiments. Indeed, some research has shown that at high transcript concentrations, the nonspecific component of the PM signal is a negligible part of the MM signal. Five variations of a method to model and estimate levels of gene expression on Affymetrix chips, without the use of MM cells, are presented. To test the validity of the algorithms, six different concentrations of human liver cRNA were prepared. Each of these solutions was then hybridized on five Affymetrix hg95A arrays using five different scanners. The expression estimates obtained using each of the five algorithms bore a strong linear relationship to the cRNA concentrations, particularly at low cRNA concentrations. The five variations were applied to a Latin square data set and the results compared with those obtained using Affymetrix MAS 5.0 and using robust multichip analysis. R² values obtained using the new techniques were comparable, while fold changes were superior.